crawl4md

crawl4md is a minimal, clean CLI tool that crawls web pages or sitemaps and converts them into structured Markdown files.

The project is intentionally designed to stay simple, deterministic, and easy to extend — without unnecessary complexity or hidden behavior.

---

Philosophy

• Minimal: only what is needed, nothing more
• Deterministic: same input → same output
• Transparent: no magic, clear processing steps
• Composable: ideal as a building block for pipelines (e.g. RAG)

---

Features

• Crawl from:
  • sitemap.xml
  • explicit page lists
• Clean Markdown output via exchangeable parser backends
• Deterministic file structure based on URL paths
• YAML-based project configuration
• CLI-first workflow (uv-compatible)
• Clear, readable progress output

---

Installation

There are two ways to use crawl4md.

Use the Batch Crawler

If you want to use the project directly for batch crawling via crawl.yml, clone the repository:

git clone git@github.com:ixnode/crawl4md.git && cd crawl4md

Then continue with the configuration section below.

Use the Python Package

If you want to build your own tooling on top of crawl4md, install it as a package:

pip install crawl4md

Or with uv:

uv add crawl4md

For local development inside the repository:

uv sync

---

Configuration

The CLI reads a crawl.yml file from the current working directory.

Create it from the example:

cp crawl.yml.example crawl.yml

Minimal example:

projects:
    planes:
        type: pages
        crawl:
            parser: kreuzberg-dev
            parse_type: markdown
        sources:
            - https://de.wikipedia.org/wiki/Boeing_707
            - https://de.wikipedia.org/wiki/Boeing_717
        preprocessing:
            markdown:
                enabled: true
                remove_html_comments: true
                normalize_whitespace: true

    pydantic:
        type: sitemap
        crawl:
            parser: kreuzberg-dev
            parse_type: markdown
        sources:
            - https://pydantic.dev/sitemap.xml
        preprocessing:
            markdown:
                enabled: false

Available project settings:

• type: pages or sitemap
• sources: list of page URLs or sitemap URLs
• profile: optional defaults such as wikipedia for crawl, normalization, and preprocessing (loaded from profiles/*.yml)
• crawl.parser: kreuzberg-dev or crawl4ai
• crawl.parse_type: markdown; markdown-fit is available with crawl4ai
• normalization.*: HTML normalization options (enabled, entities, hidden_elements, urls, references), all default to true
• preprocessing.markdown.enabled: enables Markdown cleanup
• preprocessing.markdown.*: optional cleanup rules such as ensure_h1, remove_html_comments, remove_sections, and normalize_whitespace

For the full configuration, see crawl.yml.example. For details about all Markdown preprocessing options, see docs/markdown_preprocessing.md.

---

Usage

After cloning the repository and creating crawl.yml, use:

crawl planes
crawl pydantic

Or with uv inside the project:

uv run crawl planes
uv run crawl pydantic

---

Python API

crawl4md can also be used as a Python package after installing it with pip install crawl4md or uv add crawl4md.

The public API exports two recommended default classes:

• MarkdownFetcher: fetches a URL and returns Markdown
• MarkdownConverter: converts an existing HTML string into Markdown

Both currently use the recommended kreuzberg-dev backend.

The concrete parser classes are exported too:

• MarkdownFetcherKreuzbergDev
• MarkdownConverterKreuzbergDev
• MarkdownFetcherCrawl4AI
• MarkdownConverterCrawl4AI
• ParseType
• MarkdownPreprocessingConfig

Use the default aliases unless you explicitly need a specific parser backend.

All fetchers provide:

• fetch(url): async URL fetch and Markdown conversion
• fetch_sync(url): sync URL fetch and Markdown conversion

All converters provide:

• convert(html, url=None): async HTML-to-Markdown conversion
• convert_sync(html, url=None): sync HTML-to-Markdown conversion

Common constructor arguments:

• config: a MarkdownPreprocessingConfig
• parse_type: usually "markdown"
• content_selector: optional CSS selector for selecting only part of the HTML before conversion

Configure Parse Type

When using crawl.yml, use projects.<name>.crawl.parser to choose the parser:

• "kreuzberg-dev": recommended default, supports parse_type: markdown
• "crawl4ai": supports parse_type: markdown and parse_type: markdown-fit

In Python, use the concrete class when you need a specific parser backend. Use ParseType to control how Markdown is generated:

• "markdown": raw markdown output
• "markdown-fit": cleaned and reduced markdown output via crawl4ai

Configure Preprocessing

Use MarkdownPreprocessingConfig to enable optional cleanup steps.

For the full list of preprocessing options, see docs/markdown_preprocessing.md.

Simple example:

from crawl4md import MarkdownPreprocessingConfig

config = MarkdownPreprocessingConfig(
    enabled=True,
    remove_html_comments=True,
    normalize_whitespace=True,
)

Fetch Markdown From a URL

Use MarkdownFetcher if you want to fetch a page and directly receive Markdown.

from crawl4md import MarkdownFetcher, MarkdownPreprocessingConfig

config = MarkdownPreprocessingConfig(enabled=True)
fetcher = MarkdownFetcher(config=config, parse_type="markdown")

markdown = fetcher.fetch_sync("https://example.com")
print(markdown)

Async version:

import asyncio

from crawl4md import MarkdownFetcher, MarkdownPreprocessingConfig

config = MarkdownPreprocessingConfig(enabled=True)
fetcher = MarkdownFetcher(config=config, parse_type="markdown")

markdown = asyncio.run(fetcher.fetch("https://example.com"))
print(markdown)

Convert HTML to Markdown

Use MarkdownConverter if you already have HTML and only want the conversion step.

from crawl4md import MarkdownConverter, MarkdownPreprocessingConfig

html = "<html><body><h1>Hello</h1><p>World</p></body></html>"

config = MarkdownPreprocessingConfig(enabled=True, ensure_h1=True)
converter = MarkdownConverter(config=config, parse_type="markdown")

markdown = converter.convert_sync(html=html, url="https://example.com")
print(markdown)

Async version:

import asyncio

from crawl4md import MarkdownConverter, MarkdownPreprocessingConfig

html = "<html><body><h1>Hello</h1><p>World</p></body></html>"

config = MarkdownPreprocessingConfig(enabled=True, ensure_h1=True)
converter = MarkdownConverter(config=config, parse_type="markdown")

markdown = asyncio.run(
    converter.convert(html=html, url="https://example.com")
)
print(markdown)

Limit Conversion to Part of the HTML

Use content_selector to convert only the matching HTML elements before Markdown conversion.

from crawl4md import MarkdownConverter, MarkdownPreprocessingConfig

html = """
<html>
    <body>
        <nav>Navigation</nav>
        <main><h1>Hello</h1><p>World</p></main>
    </body>
</html>
"""

converter = MarkdownConverter(
    config=MarkdownPreprocessingConfig(enabled=True),
    parse_type="markdown",
    content_selector="main",
)

markdown = converter.convert_sync(html=html, url="https://example.com")
print(markdown)

The same option is available on MarkdownFetcher.

Use a Specific Parser Backend

Use MarkdownFetcherKreuzbergDev or MarkdownConverterKreuzbergDev when you want the recommended backend explicitly.

Use MarkdownFetcherCrawl4AI or MarkdownConverterCrawl4AI when you need crawl4ai, for example parse_type="markdown-fit":

from crawl4md import MarkdownFetcherCrawl4AI, MarkdownPreprocessingConfig

fetcher = MarkdownFetcherCrawl4AI(
    config=MarkdownPreprocessingConfig(enabled=True),
    parse_type="markdown-fit",
)

markdown = fetcher.fetch_sync("https://example.com")
print(markdown)

---

Output Structure

Markdown files are stored deterministically based on the URL path:

crawled/<project>/<url-path>.md

Example:

crawled/planes/wiki/Boeing_707.md

Rules:

• Domain is ignored
• URL path is preserved
• / → index.md
• Query parameters are ignored

---

Example Output

1/2 Crawl https://de.wikipedia.org/wiki/Boeing_707
- Fetching ... done
- Processing ... done
- Writing crawled/planes/wiki/Boeing_707.md ... done

---

Use Cases

• RAG data ingestion
• Website snapshotting
• Knowledge base generation
• Offline documentation

---

Project Structure

src/crawl4md/
├─ cli.py
├─ config.py
├─ sitemap.py
├─ crawler.py
├─ paths.py
└─ writer.py

---

Notes

• No recursive crawling (by design)
• No hidden caching or transformations
• Focus on clean Markdown output only

---

License

This project is licensed under the MIT License. See the LICENSE.md file for details.

Authors

• Björn Hempel bjoern@hempel.li - Initial work - https://github.com/bjoern-hempel

---

Parser Backends

crawl4md is designed as a small orchestration layer around exchangeable HTML-to-Markdown backends.

It currently integrates the excellent crawl4ai project and html-to-markdown by kreuzberg-dev. Both libraries solve the conversion problem from different angles; crawl4md keeps the project workflow, preprocessing, path handling, and writing logic independent from the selected parser.

Why use crawl4md around these parser backends:

• project-based batch crawling via crawl.yml
• support for both page lists and sitemap-driven crawls
• deterministic output paths for generated Markdown files
• optional Markdown cleanup rules for better downstream text quality
• a small CLI and Python API focused on URL or HTML to Markdown workflows
• clearer separation between fetching, conversion, preprocessing, and writing

In short: the parser backend can change, while crawl4md keeps the surrounding crawl configuration, deterministic output, and Markdown cleanup workflow stable.

Troubleshooting

Wikipedia returns 403 Forbidden

Some websites, especially Wikimedia/Wikipedia, may block direct HTTP requests depending on the Python runtime, TLS fingerprint, request frequency, IP reputation, or server-side bot detection.

Example error:

httpx.HTTPStatusError: Client error '403 Forbidden'
Please respect our robot policy ...

This is not necessarily a crawl4md bug. The same request may work in one Python environment and fail in another.

Known workaround:

uv python install 3.14.0
uv venv --python 3.14.0
uv sync

Then run again:

uv run crawl <profile>

If the problem persists:

• reduce request frequency
• avoid repeated crawling of the same Wikimedia pages
• use a proper User-Agent
• respect Wikimedia's robot policy
• retry later